Machine learning (ML) methods have been widely explored for predicting material properties. However, due to the rapid development of ML techniques and the diversity of available models, performance comparisons between traditional and graph-based machine learning models remain limited. Therefore, we evaluate 11 conventional ML models alongside the graph neural network-based Crystal Graph Convolutional Neural Network (CGCNN) for predicting three key properties—formation energy (E_f), band gap (E_g), and energy above hull (E_h)—across a dataset comprising single perovskites, double perovskites, and their combined structures. The results demonstrate that for single perovskites, CGCNN exhibits gains of over 20% in the root mean square error (RMSE) relative to the second-best model (Gradient Boosting Regression), achieving values of 0.205 eV/atom (E_f), 0.718 eV (E_g), and 0.167 eV/atom (E_h). Prediction accuracy for double perovskites is significantly enhanced by training CGCNN on a combined dataset, particularly for E_h, where the coefficient of determination (R²) improves approximately 68.1-fold compared to models trained exclusively on double-perovskite data. Feature importance analysis via one-shot, permutation-based, and recursive feature elimination (RFE) methods reveals that optimal model performance requires retention of at least the top 20 critical features. Furthermore, feature utilization patterns of CGCNN across different prediction tasks are visualized. This work provides actionable guidelines for model selection and feature engineering in perovskite property prediction, establishing a benchmark for future ML-driven materials discovery.